paper review: high-performance transaction processing in
TRANSCRIPT
![Page 1: Paper review: High-performance Transaction Processing in](https://reader036.vdocuments.site/reader036/viewer/2022081503/62a22f070589d4247a350bd3/html5/thumbnails/1.jpg)
Paper review: High-performance Transaction Processing in SAP HANA Erfan Zamanian Feb. 2015
![Page 2: Paper review: High-performance Transaction Processing in](https://reader036.vdocuments.site/reader036/viewer/2022081503/62a22f070589d4247a350bd3/html5/thumbnails/2.jpg)
Background • SAP HANA:
• Main-memory database • Supports both analytical and transactional workloads • Allows for column and row store • Uses MVCC with distributed SI and locking scheme • The deadlock detection is handled centrally.
2
![Page 3: Paper review: High-performance Transaction Processing in](https://reader036.vdocuments.site/reader036/viewer/2022081503/62a22f070589d4247a350bd3/html5/thumbnails/3.jpg)
Main Contribution • A number of optimizations for:
• Distributed Snapshot Isolation (SI) • Exploiting locality of transactions • Local transactions pay for local coordination • Global transactions pay for global coordination
• Two-phase commit protocol • Minimizing synchronized logging • Less communication cost
Category of the Paper
Improvement over existing work�
3
![Page 4: Paper review: High-performance Transaction Processing in](https://reader036.vdocuments.site/reader036/viewer/2022081503/62a22f070589d4247a350bd3/html5/thumbnails/4.jpg)
Weaknesses • Lacks scientific methodology
• Contributions do not support all claims • Incompatibility of text and figure
• Leaves some questions unanswered
4
![Page 5: Paper review: High-performance Transaction Processing in](https://reader036.vdocuments.site/reader036/viewer/2022081503/62a22f070589d4247a350bd3/html5/thumbnails/5.jpg)
Methodology • For the distributed snapshot isolation:
• Introduced the notion of transaction token • A little bit verbose
• For 2PC optimization: • Ad hoc in nature (early commit ack, skipping writes, group 2PC) • More of an engineering effort than research
• No experiments, numbers, graphs, … • Particularly essential for this type of papers
5
![Page 6: Paper review: High-performance Transaction Processing in](https://reader036.vdocuments.site/reader036/viewer/2022081503/62a22f070589d4247a350bd3/html5/thumbnails/6.jpg)
Claims vs. Results • The paper’s main claim:
• High throughput for OLTP while allowing OLAP
• But only discusses optimization for OLTP • For example:
• Impacts of the delta buffer on OLAP? • How frequently/when merge delta with with main store? • Index on these tables?
• Bottom line: this paper does not discuss OLAP
6
![Page 7: Paper review: High-performance Transaction Processing in](https://reader036.vdocuments.site/reader036/viewer/2022081503/62a22f070589d4247a350bd3/html5/thumbnails/7.jpg)
Incompatibility of text and figure
7
![Page 8: Paper review: High-performance Transaction Processing in](https://reader036.vdocuments.site/reader036/viewer/2022081503/62a22f070589d4247a350bd3/html5/thumbnails/8.jpg)
Unanswered Questions • In section 2.2
• Each record has local_TID and global_TID • A local write transaction sees local_TID • A global write transaction sees global_TID and local_TID
• Algorithm correctness?
Record_ID Local_TID Global_TID
1 5 6
2 7 8
3 9 2
4 5 6
8
![Page 9: Paper review: High-performance Transaction Processing in](https://reader036.vdocuments.site/reader036/viewer/2022081503/62a22f070589d4247a350bd3/html5/thumbnails/9.jpg)
Unanswered Questions
• Local transactions always commit
• Paper’s assumptions: • Short-running transactions are local-only • Long-running transactions are multi-node
• SI follows “first committer wins”
Unaddressed question: How is starvation handled?
9
![Page 10: Paper review: High-performance Transaction Processing in](https://reader036.vdocuments.site/reader036/viewer/2022081503/62a22f070589d4247a350bd3/html5/thumbnails/10.jpg)
Suggestions • In HANA:
• Coordinator assigns a range of TIDs to each node • Eliminates the communication cost for TID request • Coordinator discards the unused TIDs periodically
• Suggestion: further reduce the communication cost by: • Assigning fixed TID numbering scheme (e.g. mod, hash)
10
![Page 11: Paper review: High-performance Transaction Processing in](https://reader036.vdocuments.site/reader036/viewer/2022081503/62a22f070589d4247a350bd3/html5/thumbnails/11.jpg)
Conclusion • The paper presents some novel optimizations
• Not quite ready for publication in such a venue
• Benefits from • revising some sections • improving the flow of the paper • following a systematic approach • adding experiments
11
![Page 12: Paper review: High-performance Transaction Processing in](https://reader036.vdocuments.site/reader036/viewer/2022081503/62a22f070589d4247a350bd3/html5/thumbnails/12.jpg)
And now, let’s welcome the authors
12